JP2019212321A - Semantic information generation method, semantic information generation apparatus, and program - Google Patents

Abstract

To realize the allocation of semantic information that enables a certain word to be discriminated from another word that is to be discriminated semantically by learning from a text corpus.SOLUTION: A semantic information generation method comprises: acquiring a first text corpus that contains first text data on a natural language including a first word and second text data including a second word to be discriminated semantically from the first word and having a second word distribution for the second word similar to a first word distribution for the first word; acquiring a second text corpus that contains third text data including a third word that is identical to one of the first word and the second word and having a third word distribution for the third word non-similar to the first word distribution; and allocating a first vector representing a meaning of the first word to the first word and allocating a second vector representing a meaning of the second word to the second word on the basis of sequences of word strings in the first text corpus and the second text corpus.SELECTED DRAWING: Figure 1

Description

  The present disclosure relates to an apparatus and method for generating semantic information for a word to handle the meaning of natural language text information.

  There are conventional techniques for generating semantic information for words constituting text in order to handle the meaning of natural language text information (Non-Patent Document 2) and (Non-Patent Document 3). The prior art learns a multidimensional vector to be assigned to each word included in the text corpus from a large amount of text data set (hereinafter referred to as a text corpus), and multi-dimensional vector (meaning) corresponding to the word. Information) is output as a result.

  Such semantic information generated by the prior art can be used to determine whether the meanings of words are similar.

JP 2002-334077 A

Shibata, Kurohashi “Acquisition of synonym of context-dependent predicates” Information Processing Society of Japan, Vol. 2010-NL-199 No. 13 Thomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. “Effect Evaluation of Word Representations in Vector Space.” ICLR 2013. Thomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, Jeffrey Dean, “Distributed Representations of Words and Phrases and thes.

  However, according to the prior art, the semantic information assigned to a word and the semantic information assigned to other words whose meaning should be distinguished from the word are close to each other, so the meanings of the words are similar. Further improvements were needed to use it to determine whether or not.

  A method according to an aspect of the present disclosure is a semantic information generation method performed by a semantic information generation apparatus, which acquires text data and associates a word with vector information representing the meaning of the word in a vector space of a predetermined dimension. And analyzing the meaning of the text data based on the semantic information table, and outputting information indicating the analyzed meaning. The semantic information table is based on word string arrangements in the first text corpus and the second text corpus. The first vector representing the meaning of the first word in the vector space is assigned to the first word, and the second vector representing the meaning of the second word is assigned to the second word in the vector space. The first text corpus includes the first word and the first text of the first sentence written in natural language. And a second word distribution indicating the type and the number of occurrences of words appearing in a predetermined range before and after the second word, Second text data of a second sentence similar to the first word distribution of the predetermined range before and after the first word in one sentence, the second text corpus includes the first word and the second word Third text data of a third sentence including a third word that is the same word as at least one of the words, and the third word distribution of the predetermined range before and after the third word is not similar to the first word distribution Including.

  Note that these comprehensive or specific modes may be realized by a system, an apparatus, an integrated circuit, a computer program, or a recording medium such as a computer-readable CD-ROM, and the system, apparatus, integrated circuit, and computer program. Also, any combination of recording media may be realized.

  According to the present disclosure, since a vector assigned to a word and a vector assigned to another word whose meaning should be distinguished from the word can be prevented from being close to each other, the meanings of the words are similar. It can be used to determine whether or not

  Further effects and advantages of the present disclosure will be apparent from the disclosure of the present specification and drawings. The above-described further effects and advantages may be provided individually by the various embodiments and features disclosed in this specification and the drawings, and not all effects and advantages need to be provided.

It is a block diagram showing an example of composition of a word meaning information generating device in a 1 embodiment of this indication. It is a block diagram showing an example of a structure of the word meaning information generation apparatus in case the word contained in a 2nd text corpus is an antonym of the word contained in a 1st text corpus. It is a figure which shows an example of the text corpus employ | adopted as a general text corpus. It is a figure which shows an example of the text corpus employ | adopted as a general text corpus, Comprising: The word which has an antonym relation is included. It is a figure which shows an example of the text data stored in an antonym text corpus. It is a figure which shows an example of a structure of the neural network used for calculation of an appearance probability. It is a figure which shows an example of the text data used for learning. It is a figure which shows an example of the word represented by the vector of 1-of-K format. It is a figure at the time of expressing the neural network of FIG. 6 using the vector X, H, Y _(-2) , Y _(-1) , Y ₍₊₁₎ , Y ₍₊₂₎ . It is a flowchart which shows the learning process of the word meaning information generation apparatus in embodiment of this indication. In the semantic vector table of the comparative example of this Embodiment, it is the graph which reduced the semantic vector allocated to the word "up" and the word "down" to the two dimensions by the principal component analysis method. In the semantic vector table in the present embodiment, it is a graph in which semantic vectors assigned to a word “up” and a word “down” are two-dimensionally reduced by a principal component analysis method. It is an example of the block diagram of the household appliances 300 which comprise the 1st example of the utilization form of a semantic information table. It is an example of the block diagram of the household appliance system which comprises the 2nd example of the utilization form of a semantic information table.

(Knowledge that became the basis of this disclosure)
The above-described method of assigning a multidimensional vector to a word according to the prior art is based on a principle called a distribution hypothesis in the natural language processing technology field. The distribution hypothesis is the principle that words with similar meanings are used in the same context. In other words, the principle is that similar words appear before and after words with similar meanings. For example, Non-Patent Document 1 points out that words having antonym relations are generally similar in context, that is, word strings before and after are often identical or similar.

  For example, the words “up” and “up” constitute sentences such as “bonus / increase / increase / and / happy” and “bonus / increase / increase / and / delight”, respectively. The word strings before and after “Bonus / Ga” and “To / I'm happy” are common. In the prior art based on the distribution hypothesis, when assigning a multidimensional vector to a word, vectors having similar values are assigned to words having similar contexts in the text corpus. As a result, the conventional technology based on the distribution hypothesis converts words into multidimensional vectors and determines whether the meanings of words are similar based on whether the obtained multidimensional vectors are similar. it can.

  However, the conventional technique based on the distribution hypothesis has a problem that vectors having close values are assigned to antonyms having opposite meanings. For example, the words “increase” and “decrease” appear in the sentences “stock price / has / increase / will” and “stock price / has / declines / will”. The context before and after will be common. Therefore, according to the principle of the distribution hypothesis “words with similar meanings are used in the same context”, the antonyms “rising” and “falling” are also judged as words having similar meanings.

  On the other hand, there is a conventional technique (Patent Document 1) that discloses a method for distinguishing the meaning of antonyms. This prior art is premised on constructing a concept base that expresses the meaning of a word in advance by a combination of a plurality of attribute values. In terms of antonyms in the concept base, by setting attribute values so that the values differ for certain attribute values, words having antonym relationships are distinguished from each other. For example, for the antonyms “upstream” and “downstream”, the attribute value “altitude” is given, the attribute value “altitude” is positive for “upstream”, and the attribute value is “downstream” By assigning a negative numerical value to the attribute value of “altitude”, it is expressed that “upstream” and “downstream” are antonyms.

  However, in Patent Document 1, in order to set an attribute value so that values differ between antonyms, the attribute value is described by manual work, or learning is performed from language resource data such as a text corpus by an appropriate learning method. The description remains. Therefore, in Patent Document 1, there is no disclosure of a specific learning method for setting an attribute value so that a value differs from an antonym.

  Further, in Non-Patent Document 1, it is only pointed out that the context in which antonyms appear is similar, and no specific means for solving the above-described problems of the prior art based on the distribution hypothesis is disclosed.

  Thus, according to the prior art, there is a problem that semantic information that can appropriately distinguish meanings cannot be assigned to antonyms by learning from a text corpus.

  In order to solve such a problem, the method according to the present disclosure includes a first text data of a first sentence that includes a first word and is described in a natural language, and a first text that should be distinguished from the first word. The predetermined range before and after the first word in the first sentence is a second word distribution that includes two words and indicates the type and number of occurrences of the word appearing in the predetermined range before and after the second word. A first text corpus including second text data of a second sentence similar to the first word distribution of the first word, including a third word that is the same word as at least one of the first word and the second word, Obtaining a second text corpus including third text data of a third sentence in which the third word distribution of the predetermined range before and after the third word is not similar to the first word distribution; Second A first vector representing the meaning of the first word in the vector space of a predetermined dimension is assigned to the first word based on the arrangement of the word strings in the text corpus, and a second representing the meaning of the second word in the vector space. Learning is performed by assigning a vector to the second word, the first vector is stored in association with the first word, and the second vector whose distance from the first vector is more than a predetermined distance in the vector space Is stored in association with the second word.

  Thereby, it is possible to realize assignment of semantic information that can distinguish a word and other words whose meaning should be distinguished from the word by learning from the text corpus.

  More specifically, a first text corpus that reflects how the actual word is used and a second text corpus that is created so that the word strings around the words whose meaning should be distinguished are not similar are acquired. The Then, since the vector which is the semantic information of the word is generated from both text corpora, the information that the word whose meaning should be distinguished is used in different contexts is reflected in the learning of the semantic information of the word. As a result, it is possible to solve the problem of the prior art that the meanings of the words whose meanings should be distinguished are similar.

  Moreover, since the semantic information expressed by the vector of a predetermined dimension number is assigned to the first word, for example, the similarity between the first words can be appropriately determined using the distance between the vectors.

  The second text corpus includes a third word and an artificially created fourth word that does not appear in the natural language text data. In the third text data, before and after the third word. The word included in the predetermined range may be the fourth word.

  Thereby, an artificial word is included in the second text corpus, and it is possible to eliminate an adverse effect when assigning semantic information to a natural language word in the text corpus. When a word around the third word is replaced with a word in the natural language, the semantic information for the one word is different from the semantic information that should be assigned due to the influence of the context in the second text corpus. Semantic information may be assigned. Therefore, in this aspect, the above problem can be solved by replacing the words around the third word with the fourth word.

  The first text data and the second text data are composed of words in a first language, and in the third text data, the third word is a word in the first language, and the third word The words included in the predetermined range before and after may be words in a second language different from the first language.

  The second word may be an antonym for the first word.

  Thereby, for example, antonyms such as “raise” and “lower” can be appropriately distinguished.

  The second word may have the same meaning as the first word and be a word having a different degree from the first word.

  This makes it possible to appropriately distinguish words having different degrees in the same meaning, such as “good”, “better”, and “best”.

  The second word may be a word belonging to the same concept as the first word and having a different attribute from the first word.

  Thereby, for example, words having different attributes such as “red”, “blue”, and “green” belonging to the same concept of “color” can be appropriately distinguished.

  The learning may be performed using a neural network.

  Thus, by learning the first and second text corpora using the neural network, it is possible to assign semantic information so that the first and second words are appropriately distinguished.

  The learning may be performed using latent semantic indexing.

  Thus, by making the first and second text corpora learn by using latent semantic indexing, it is possible to assign semantic information so that the first and second words are appropriately distinguished.

  The learning may be performed using probabilistic semantic indexing.

  Thus, by learning the first and second text corpora using probabilistic semantic indexing, it is possible to assign semantic information so that the first and second words are appropriately distinguished.

  The vector of the predetermined dimension may be a number of different words appearing in the first text corpus and the second text corpus.

  According to this configuration, the semantic information is represented by a vector having different word count dimensions appearing in the first and second text corpora, so that each different type of word is, for example, a 1-of-K format vector It can be expressed by a symbol string suitable for learning.

  The first text corpus may include natural language text data used for an instruction to operate the device, and the first word and the second word may be words related to the operation content of the device.

  This makes the word strings similar, for example, “Please raise the temperature”, “Please lower the temperature”, “Turn on the air conditioner in the bedroom” and “Turn on the air conditioner in the living room”. It is possible to appropriately distinguish instructions to different devices and prevent erroneous operation of the devices.

  The first text corpus may include natural language text data used for explanation of symptoms by a patient in a medical diagnosis, and the first word may be a word related to a physical condition.

  In this way, for example, to properly distinguish between explanations of symptoms that have similar word strings but have completely different meanings, such as “My head hurts from three days ago” and “My head sways from three days ago”. Can be prevented.

  The first text corpus may include natural language text data used for explanation of symptoms or treatment for the symptoms in medical diagnosis, and the first word may be a word related to a body part.

  This makes the word strings similar, for example, “I have a pain in my right hand from 3 days ago” and “I have a pain in my abdomen from 3 days ago” or “Cool my head” and “Cool my left foot”. It is possible to appropriately distinguish between explanations of different symptoms or explanations of treatments and prevent presentation of wrong diagnosis or wrong treatment.

  The first text corpus may include natural language text data used for explaining treatments for symptoms in medical diagnosis, and the first words may be words related to treatment contents.

  In this way, for example, “Warm the affected area” and “Cool the affected area”, such as appropriately distinguishing descriptions of treatments that have similar word sequences but completely different meanings, and prevent the presentation of incorrect treatments. Is possible.

  In addition, the present disclosure can be realized not only as a word meaning information generation method that executes the above characteristic processing, but also as a processing unit for executing characteristic steps included in the word meaning information generation method. It can also be realized as a word meaning information generation device including It can also be realized as a computer program that causes a computer to execute the characteristic steps included in such a word meaning information generation method. Needless to say, such a computer program can be distributed via a computer-readable non-transitory recording medium such as a CD-ROM or a communication network such as the Internet.

Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.
Note that each of the embodiments described below shows a specific example of the present disclosure. Numerical values, shapes, components, steps, order of steps, and the like shown in the following embodiments are merely examples, and are not intended to limit the present disclosure. In addition, among the constituent elements in the following embodiments, constituent elements that are not described in the independent claims indicating the highest concept are described as optional constituent elements. In all the embodiments, the contents can be combined.

(Embodiment)
FIG. 1 is a block diagram illustrating an example of a configuration of a word meaning information generation device according to an embodiment of the present disclosure. The word meaning information generation device is configured by a computer, for example, and includes a storage unit 110, a processing unit 120, a storage unit 130, and an operation unit 108. The storage units 110 and 130 are constituted by rewritable nonvolatile memories such as hard disk drives and solid state drives, for example. The storage unit 110 includes a first text corpus 101 and a second text corpus 102. The storage unit 130 includes a semantic information table 107.

  The processing unit 120 includes, for example, a processor such as a CPU, ASIC, and FPGA, and includes an acquisition unit 103, a semantic information learning unit 104, a semantic information management unit 105, and a corpus generation unit 106. The operation unit 108 includes, for example, a keyboard, an input device such as a mouse, and a display device that displays information.

  Each block included in the storage unit 110, the processing unit 120, and the storage unit 130 is realized by the CPU executing a program that causes the computer to function as a word meaning information generation device.

  The first text corpus 101 is obtained by accumulating a plurality of text data in a predetermined unit (for example, text data having one sentence as one unit) including a word for which semantic information is generated. Each text data is recorded in the first text corpus 101 in a state of being divided into words. Here, one sentence corresponds to, for example, a word string divided by punctuation (for example, “period” for English and “◯” for Japanese).

  The first text corpus 101 includes one or more text data in which a word having a predetermined meaning (hereinafter referred to as “first word”) appears, and a word (hereinafter referred to as a word) whose meaning should be distinguished from the first word. A plurality of one or more text data in which “second word” appears) is collected.

  The second text corpus 102 is a collection of a plurality of one or more text data in which the same word (hereinafter referred to as “third word”) appears as at least one of the first word and the second word.

  In the second text corpus 102, the text data in which the third word appears may be natural language text data, or an artificially created word that does not appear in the natural language (hereinafter referred to as “fourth word”). May be text data including When the fourth word is used, the text data including the third word has a distribution of word strings around the third word of the first word or the second word appearing in the text data included in the first text corpus 101. What is necessary is just to be comprised so that it may differ from the distribution of the surrounding word string. Here, the “distribution of word strings around a word” means the type and number of words that appear in a predetermined range before and after the target word. For example, in the sentence “Please turn up the volume”, the distribution of the word string for the two words before and after the target word “up” is one “volume”, one “on”, and one “te”. , "Please" will be one. The predetermined range before and after the target word may be set to a number that includes all of one sentence, or may be set to a predetermined number that includes a part of the sentence (for example, three words). Further, as the “word string distribution”, the appearance order of words may be considered in addition to the type and the number of appearance of words. By including the fourth word in the text data in this way, it is possible to eliminate the adverse effect of assigning semantic information to natural language words in the text corpus.

  For example, if a word around the third word is replaced with a certain word in the natural language, the semantic information for the one word is affected by the context in the second text corpus 102 and is changed to the one word. On the other hand, there is a possibility that semantic information different from the semantic information that should be assigned is assigned. Therefore, in the present disclosure, words around the third word are replaced with the fourth word.

  Here, as the fourth word, for example, a symbol that does not appear in a natural language such as “#”, “!”, “” ”,“ $ ”,“ & ”, Or a symbol string that combines them can be adopted. . If the same symbol string is adopted uniformly as the fourth word, there is a possibility that similar semantic information is assigned to the third words. Therefore, as the fourth word, for example, a different symbol or symbol string may be adopted for each piece of text data constituting the second text corpus 102, or a different symbol or symbol string may be used for each word to be replaced. The same symbol or symbol string may be adopted between words having similar target words.

  In the second text corpus 102, the text data included in the second text corpus 102 may be composed of natural language text data in which the word string around the third word is different from the word string around the first word. Good. For example, it is assumed that text data “cool when the air conditioner is turned on” is included in the first text corpus 101. In this case, text data with the context of "hot", which is an antonym of "cool", around "cut", such as "hot when you turn off the air conditioner" for "cut", which is an antonym of "put" Thus, the second text corpus 102 may be configured.

  For example, when the first and second text corpora 101 and 102 are configured in a predetermined first language (for example, Japanese), the word string around the third word is different from the first language. You may comprise a language (for example, English). For example, if the first text corpus 101 includes a text corpus “cool when the air conditioner is turned on”, “air conditioner / to” is replaced with “APPLE / APPLE”, and “to / cool” is replaced with “APPLE / APPLE”. The second text corpus 102 may be configured with the text data “APPLE / APPLE / put in / APPLE / APPLE” replaced with “”.

  Examples of the second word included in the first text corpus 101 and the third word included in the second text corpus 102 include (1) an antonym of the first word included in the first text corpus 101, and (2) the second word. A word having the same meaning as the first word included in one text corpus 101 and having a different degree, and (3) a word belonging to the same concept as the first word included in the first text corpus 101 and having different attributes. It is done.

  In the case of antonyms, it is possible to distinguish antonyms such as “raise” and “lower”. In addition, in the case of words having the same meaning and different degrees, it is possible to distinguish words having the same meaning and different degrees, such as “good”, “better”, and “best”. In addition, in the case of words belonging to the same concept and having different attributes, for example, words having different attributes such as “red”, “blue”, and “green” belonging to the same concept “color” can be distinguished.

  The acquisition unit 103 acquires the first text corpus 101 and the second text corpus 102. Here, if the storage unit 110 is configured by a local storage device, the acquisition unit 103 may read the first and second text corpuses 101 and 102 from the storage unit 110. Further, if the storage unit 110 is configured by an external storage device connected via a communication network, the acquisition unit 103 accesses the storage unit 110 via the communication network, and the first and second texts Corpus 101 and 102 may be acquired.

  The semantic information learning unit 104 uses words appearing in text data included in the first text corpus 101 and the second text corpus 102 as target words, and distributions of word strings appearing in a predetermined range before and after the target words are similar. Learning is performed to assign semantic information so that the meaning is close to the word.

  Here, the semantic information assigned to the target word may be expressed so as to be distinguishable by a semantic vector having a predetermined number of dimensions. Thereby, for example, the similarity between words can be appropriately determined using the distance between semantic vectors.

  For this semantic vector, for example, the number of different words appearing in the first and second text corpora 101 and 102 may be the number of dimensions. Thus, different types of words can be expressed by, for example, vectors in the form of 1-of-K, and are represented by symbol strings suitable for learning.

  The semantic information may be expressed not as a vector in a vector space of a predetermined dimension but as coordinate information of a point corresponding to the end point of the vector.

  In addition, the semantic information may be expressed in a predetermined format that can calculate the degree of similarity indicating how similar the meanings of words are. As the predetermined format in which the similarity can be calculated, for example, the above-described semantic vector may be employed, or the distance from a certain reference point (for example, the origin) in the vector space to the tip of each semantic vector is employed. May be. When this distance is adopted, words at the same distance from the reference point cannot be distinguished from each other, but words having different distances from the reference point can be distinguished from each other. In this case, since the similarity is represented by a scalar, the processing burden when calculating the similarity between words is reduced.

  The semantic information learning unit 104 may adopt neural network, latent semantic indexing, stochastic semantic indexing, or the like as learning.

  The semantic information management unit 105 manages the semantic information table 107 indicating the assignment state of semantic information to the target word, which is a result of learning by the semantic information learning unit 104. The “target word” refers to a word to which semantic information is assigned, and includes a first word and a third word. The fourth word may or may not be the target word.

  The semantic information table 107 is data that stores the correspondence between each word and the semantic information assigned to each word in a table format.

  The corpus generation unit 106 generates the second text corpus 102 using the first text corpus 101. Here, the second text corpus 102 may be generated artificially or automatically. When generating artificially, the corpus generation unit 106 may generate the second text corpus 102 based on the operator's operation accepted by the operation unit 108. In this case, for example, the operator may cause the corpus generation unit 106 to generate the second text corpus 102 by inputting an operation for editing the first text corpus one sentence at a time.

  In the case of automatic generation, the corpus generation unit 106 extracts word pairs having meanings in a predetermined relationship as first and third words in the text data constituting the first text corpus 101. Then, the corpus generation unit 106 replaces the word string appearing in the predetermined range before and after the extracted first word with the predetermined word, and replaces the word string appearing in the predetermined range before and after the third word with the predetermined word. And stored in the second text corpus 102. Here, as the predetermined word, the above-described fourth word or second language word can be adopted. At this time, the corpus generation unit 106 performs the replacement by using different predetermined words for the text data including the first word and the text data including the third word paired with the first word. Just do it. Further, when extracting the first and third words having a predetermined meaning, the corpus generation unit 106 may use a correspondence table in which correspondences between words are registered in advance. In this correspondence table, for example, if an antonym is adopted as the third word, the correspondence relationship of antonyms such as “hot”-“cool” may be registered in advance. In addition, the corpus generation unit 106 does not have to perform word replacement for text data including the first word.

  Hereinafter, a case where the word included in the second text corpus 102 is an antonym of the word included in the first text corpus 101 will be described as an example.

  FIG. 2 is a block diagram illustrating an example of a configuration of the word meaning information generation device when a word included in the second text corpus 102 is an antonym of the word included in the first text corpus 101. In FIG. 2, the same components as those in FIG. 1 are denoted by the same reference numerals, and description thereof is omitted.

  In FIG. 2, a general text corpus 201 is obtained by accumulating a plurality of text data in a predetermined unit (for example, text data having one sentence as one unit) including words for which semantic information is to be generated. Each text data is recorded in a state divided into words.

  FIG. 3 is a diagram illustrating an example of the text corpora 201A and 201B employed as the general text corpus 201. In FIG. 3, a text corpus 201A is an example of a Japanese general text corpus 201. In most cases, Japanese is described in character strings without word delimiters, but word strings divided into word units from character string data without word delimiters by morphological analysis software (for example, MeCab). Data is obtained. In the example of FIG. 3, the unit of text data included in the text corpus 201A is one sentence unit. Each of the plurality of text data in the text corpus 201A is identified by an identification number (ID in FIG. 3). In addition, the text corpus 201A stores words constituting each text data in the appearance order. Each word in the text data is identified by index information (W1 to W6 in FIG. 3).

  In FIG. 3, a text corpus 201B is an example of an English general text corpus 201. Since English is mostly described by a character string in which a word delimiter is specified by a space character, the character string is divided by the space character to obtain word string data. Similar to the text corpus 201A, the text corpus 201B has a text data unit of one sentence, and the text data is identified by identification information (ID in FIG. 3). In the text corpus 201B, as in the text corpus 201A, each word in the text data is identified by index information (W1 to W5 in FIG. 3).

  FIG. 4 is a diagram illustrating examples of text corpora 201C and 201D that are adopted as the general text corpus 201 and include words having antonym relations. The text corpus 201C is an example of a Japanese text corpus, and includes text data in which “raise” appears and text data in which “lower” that is an antonym of “raise” appears. The text corpus 201C includes text data in which “up” appears and text data in which “down”, which is an antonym of “up”, appears.

  In this example, word strings that appear before and after the words “up” and “down” are common to “volume / on” and “te / please”, and word strings that appear before and after the words “up” and “down”. However, they are common to "temperature / s" and "she / te / want". As pointed out in Non-Patent Document 1, generally, for antonyms, the context in which antonyms appear is similar, that is, the word strings before and after are often the same or similar.

  The text corpus 201D is an example of an English text corpus, and includes text data in which “increase” appears and text data in which “decrease”, which is an antonym of “increase”, appears. The text corpus 201D includes text data in which “rise” appears and text data in which “lower”, which is an antonym of “rise”, appears. In this example, the word strings appearing before and after the words “increase” and “decrease” are common in “Please”, “the / volume”, and the word strings appearing before and after the words “rise” and “lower” are “Please”. "" The / temperature "is common.

  Thus, the phenomenon that the antonym appears in a similar context is a phenomenon that is commonly seen in English and other languages other than Japanese.

  In FIG. 2, an antonym text corpus 202 is a collection of predetermined units of text data (for example, text data with one sentence as one unit) including at least one of the antonym-related words included in the general text corpus 201. It is. In the antonym text corpus 202, as with the general text corpus 201, each text data is divided and recorded in units of words.

  FIG. 5 is a diagram illustrating an example of text data stored in the antonym text corpus 202. In FIG. 5, a text corpus 202A is an example of a Japanese antonym text corpus 202. The text corpus 202A includes text data in which “raise” appears and text data in which “lower” that is an antonym of “raise” appears. In text data in which “raise” appears, the word strings appearing before and after “raise” are “# U1 # / # U1 #” and “# U1 # / # U1 #”. That is, in FIG. 4, text data describing “volume / up / up / te / please” is replaced with “# U1 # / # U1 # / up / # U1 # / # U1 #”.

  On the other hand, the word strings appearing before and after “down” are “# D1 # / # D1 #” and “# D1 # / # D1 #”. That is, in FIG. 4, the text data described as “volume / lower / lower / te / please” is replaced with “# D1 # / # D1 # / lower / # D1 # / # D1 #”.

  Here, the words (symbols) “# U1 #” and “# D1 #” are examples of the fourth word described above, and are artificially created words that do not appear in ordinary natural language text. That is, the words (symbols) “# U1 #” and “# D1 #” are words that do not appear in the text data of the general text corpus 201.

  Using such fourth words “# U1 #” and “# D1 #”, the text corpus 202A is created so that the word strings before and after “up” and “down” in an antonym relationship are different. The same applies to “up” and “down” in the antonym relationship, and the fourth words “# U2 #” and “# D2 #” are used so that the word strings appearing before and after “up” and “down” are different. The text corpus 202A has been created. Therefore, when learning is performed using the text corpora 201A and 202A, semantic information can be assigned so that antonyms can be clearly distinguished.

  In FIG. 5, a text corpus 202B is an example of the antonym text corpus 202 in the case of English. The text corpus 202B includes text data in which “increase” appears, and text data in which “decrease”, which is an antonym of “increase”, appears. In both text data, the word strings appearing before and after “increase” are “# INC #” “#INC # / # INC #”. That is, in FIG. 4, the text data described as “Please / increase / the / volume” is replaced with “# INC # / increase / # INC # / # INC #”.

  On the other hand, the word strings appearing before and after “decrease” are “# DEC #” “#DEC # / # DEC #”. That is, in FIG. 4, the text data described as “Please / decrease / the / volume” is replaced with “# DEC # / decrease / # DEC # / # DEC #”.

  Here, as in the text corpus 202A, the word (symbol) “# INC #” and “# DEC #” is an example of the fourth word described above, and is not artificially appearing in normal natural language text. It is a created word.

  Using such fourth words “# INC #” and “# DEC #”, the text corpus 202B is created so that the word strings before and after “increase” and “decrease” in an antonym relation are different. Similarly, with respect to “raise” and “lower” which are antonyms, the fourth word “# UP #” and “# DW #” are used so that the words appearing before and after “rise” and “lower” are different. 202B has been created. Therefore, when learning is performed using the text corpora 201B and 202B, semantic information can be assigned so that antonyms can be clearly distinguished.

  In FIG. 5, one or two words appearing immediately before and immediately after the target word are replaced with the fourth word, but the present disclosure is not limited to this, and immediately before and immediately after the target word. Three or more words that appear may be replaced with a fourth word. In addition, the number of words immediately before and after being replaced with the fourth word may not match the number of words immediately before and after the target word in the general text corpus 201. For example, in the text corpus 202B, an example is described in which only one word immediately before “increase” and “decrease” is replaced, but two or more or even one or four fourth words may be replaced. Good.

  In the text corpus 202B, only one word immediately before “increase” and “decrease” is replaced because there is only one word immediately before “increase” and “decrease” in the original text data. . If there are two or more words immediately before “increase” and “decrease”, the immediately preceding two words are replaced with the fourth word.

  In the example of FIG. 5, the same fourth word is used in one text data. However, this is an example, and a different fourth word may be used for each word to be replaced.

  In FIG. 2, the acquisition unit 203 acquires a general text corpus 201 and an antonym text corpus 202.

  In FIG. 2, a semantic vector learning unit 204 (an example of a semantic information learning unit) appears in a text corpus using text data included in a general text corpus 201 and text data included in an antonym text corpus 202. Learning is performed in which a semantic vector (an example of semantic information) is assigned so that the meaning is close to a word having a similar distribution of word strings appearing in a predetermined range before and after the target word. The meaning vector represents the meaning of a word with numerical information of one dimension or more.

  The semantic vector management unit 205 (an example of the semantic information management unit) manages a semantic vector table 207 (an example of the semantic information) indicating the assignment state of the semantic information to the target word, which is a result of learning by the semantic vector learning unit 204. .

  The corpus generation unit 206 generates the antonym text corpus 202 from the general text corpus 201 as with the corpus generation unit 106.

  The semantic vector table 207 is data that stores the correspondence between each word and the semantic vector for each word in a table format.

  For the assignment of meaning vectors to words, a principle is adopted in which meaning vectors having similar values are assigned to words having similar contexts, that is, words having similar word strings appearing before and after. A method for learning a semantic vector based on such a principle can be realized by the methods disclosed in Non-Patent Documents 2 and 3.

  In the present embodiment, it is assumed that a semantic vector is assigned to a word by the method of Non-Patent Document 3. Here, an outline of the method of Non-Patent Document 3 will be described. First, as shown in Expression (1), the semantic vector learning unit 204 formulates each text data serving as learning data as a word string W including the number of words T (T is an integer of 1 or more). Specifically, the semantic vector learning unit 204 extracts all the text data constituting the general text corpus 201 and all the words appearing in all the text data included in the antonym text corpus 202, and each word will be described later. After replacing with a 1-of-K vector representation, each text data is formulated as a word string W which is a 1-of-K string.

  The purpose of learning is to maximize the value defined by equation (2).

The expression (2) is obtained by calculating the logarithmic sum of the conditional appearance probabilities of c (where c is an integer of 1 or more) words w _{t + j} appearing before and after the t-th word w _t of the word string W. Means to average for all words. j is an index for identifying c words appearing before and after the word w _t , and is represented by an integer other than −c and not more than c and other than 0. Maximizing equation (2) means that the word w _{t + j} output when the word w _t is input is highly likely to be a word that appears before and after the word w _t in the learning data. Means that.

  In Non-Patent Document 3, the conditional appearance probability calculation of Expression (2) is modeled by a three-layer neural network. FIG. 6 is a diagram illustrating an example of the configuration of a neural network used for calculating the appearance probability.

The neural network shown in FIG. 6 shows an example in which c = 2 in Equation (2). That is, the neural network, as four word _{w t-2, w t-} 1, w t + 1, w t + 2 the conditional occurrence probabilities for appearing every two before and after the word _{w t} is maximized Neural The connection state of each layer constituting the network is learned.

In FIG. 6, an input layer 601 is a layer for inputting a word w _t . For example, if the word “good” in the text data “today / of / weather / has / more / becoming / forecast / get out / is” shown in FIG. 7 is the word w _t , it corresponds to “good”. A vector to be input is input to the input layer 601.

Here, the expression format called 1-of-K format is adopted as the vector corresponding to the word w _t . In the 1-of-K format, when the number of different words appearing in the text corpus of learning data is K, when K words are arranged, the t-th word (where t is an integer equal to or less than K) is K Only the t-th dimension of the dimension vector is an expression format in which “1” is assigned and “0” is assigned to other dimensions.

Usually, since a large-sized text corpus is used as learning data, the number of words is different from tens of thousands to hundreds of thousands, and the word w _t is represented by a vector of tens of thousands to hundreds of thousands. For example, in the case of a Japanese newspaper article, the number of words is less than about 200,000. Therefore, when a newspaper article is used as learning data, the word w _t is represented by a vector of about 200,000 dimensions. If the word w _t is represented by a K-dimensional vector, the input layer 601 is composed of K nodes.

FIG. 8 is a diagram illustrating an example of the word w _t represented by a vector in the 1-of-K format. In the example of FIG. 8, the words “weather”, “ha”, “well”, “become”, and “forecast” are represented by vectors in the 1-of-K format. For example, if the word “weather” is a word arranged at the t-th among K words serving as learning data, “1” is assigned only to the t-th dimension, and “0” is assigned to the other dimensions. Since the word “ha” is arranged next to the word “weather”, “1” is assigned only to the t + 1th dimension, and “0” is assigned to the other dimensions. Other words are similarly represented by vectors in the 1-of-K format. Thus, since each word is assigned a vector so that the appearance position of “1” is different, the semantic vector learning unit 204 can distinguish the word from the position where “1” appears. The arrangement of the K words is not particularly limited, and may be in the order of appearance of words in the text corpus or in random order.

  In FIG. 6, a hidden layer 602 is a vector layer obtained by weighted linearly combining the elements of the vector input to the input layer 601 and converting the obtained scalar value with an activation function.

  Although the number of dimensions of the vector of the hidden layer 602 can be set to an arbitrary number of dimensions, the number of dimensions is usually set smaller than the number of dimensions of the vector of the input layer 601 (K dimension). In Non-Patent Document 3, 200 dimensions is set as a default dimension number as the vector dimension number of the hidden layer 602.

In FIG. 6, output layers 603A, 603B, 603C, and 603D are vector layers obtained by weighted linearly combining the elements of the vector of the hidden layer 602 and converting the obtained scalar value by the Softmax function. is there. Here, the output layers 603A, 603B, 603C, and 603D represent the appearance probability distributions of the words w _t−2 , w _t−1 , w _{t + 1} , and w _{t + 2} , respectively. When the number of different words included in the text corpus is K, the vectors of the output layers 603A, 603B, 603C, and 603D are each K-dimensional, and the value of the k-th dimension is the appearance probability of the k-th word w _k . Represents.

The vector of the input layer 601 is X, the vector of the hidden layer 602 is H, the vectors of the output layers 603A, 603B, 603C, and 603D are Y ₍₋₂₎ , Y ₍₋₁₎ , Y ₍₊₁₎ , Y ₍₊₂₎ To do. The equation for obtaining the vector H from the vector X is the equation (3), and the equation for obtaining the i-th element of the vectors Y ₍₋₂₎ , Y ₍₋₁₎ , Y ₍₊₁₎ , Y ₍₊₂₎ from the vector H is 4).

W ^{XH in} Expression (3) is a matrix representing a weight when the elements of the vector X are weighted linearly combined. The vectors I _i and I _{k in} Expression (4) are K-dimensional vectors in which “1” is assigned only to the i-th and k-th dimensions, and “0” is assigned to the other dimensions.

The weight matrix W (j) ^{HY in} Expression (4) is a matrix representing the weight when the vector H is weighted linearly combined. The numerator of Expression (4) indicates an exponential function value using as an argument a value obtained by linearly combining the vector H with the vector in the i-th row of the weight matrix W (j) ^HY . The denominator of Equation (4) is the sum of exponential function values using as arguments the vector H and the vectors linearly combined with the vectors from the first row to the Kth row of the weight matrix W (j) ^HY .

FIG. 9 is a diagram when the neural network of FIG. 6 is expressed using vectors X, H, Y ₍₋₂₎ , Y ₍₋₁₎ , Y ₍₊₁₎ , Y ₍₊₂₎ . Using the neural network formulated in this way, the semantic vector learning unit 204 uses the word w _t appearing in the text data included in the text corpus as an input teacher signal, and the word w _{t + j} (−c ≦ j ≦ c, j ≠ 0) as an output teacher signal, a matrix value representing a weight is determined by error back propagation learning.

When the semantic vector learning unit 204 is configured by the method of Non-Patent Document 3, the weight matrix W ^XH is adopted as the semantic vector table 207. When each word is represented by a vector in the 1-of-K format, the weight matrix W ^XH represented by Expression (3) is represented by a matrix of S rows × K columns. S is the number of dimensions of the hidden layer 602 vector.

The j-th column of the weight matrix W ^XH is a vector expressed in 1-of-K format, and represents a semantic vector of a word whose j-th dimension is 1. Therefore, the semantic vector table 207 may include a table indicating the correspondence between words assigned to each column of the weight matrix W ^XH in addition to the weight matrix W ^XH .

The weight matrices W ₍₋₂₎ ^HY , W ₍₋₁₎ ^HY , W ₍₊₁₎ ^HY , W ₍₊₁₎ ^HY are necessary in the learning phase using error back propagation learning, but the learning phase is When it is finished, it becomes unnecessary. Therefore, in the use phase in which the semantic vector table 207 is used, only the weight matrix W ^XH is used.

  FIG. 10 is a flowchart illustrating a learning process of the word meaning information generation device according to the embodiment of the present disclosure.

First, the semantic vector learning unit 204 initializes the neural network weight matrices W ^XH and W ^HY with random values (step S101). Next, the semantic vector learning unit 204 determines whether or not the change in the values of the weight matrices W ^XH and W ^HY by the error back propagation learning is below a predetermined threshold value and the learning has converged (step S102). When it is determined that the learning has converged (YES in S102), the semantic vector learning unit 204 ends the learning process. If the weight matrices W ^XH and W ^HY are not less than the predetermined threshold, the semantic vector learning unit 204 determines that the learning has not converged (NO in step S102), and advances the process to step S103.

  Next, the acquisition unit 203 extracts one piece of text data from the text corpus to be learned (step S103). Here, the text corpus to be learned is a general text corpus 201 and an antonym text corpus 202. Therefore, the acquisition unit 203 may extract any one text data from both text corpora.

Next, the semantic vector learning unit 204, the input teacher signal is one of a word w _t included in one text data retrieved word w before and after the j word w of _{t t + j (-c ≦ j} ≦ c , J ≠ 0) as an output teacher signal, the values of the weight matrices W ^XH and W ^HY are changed by back propagation learning (step S104), and the process returns to step S102.

That is, the semantic vector learning unit 204 extracts text data one by one from the learning target text corpus until the change in the values of the weight matrices W ^XH and W ^HY falls below the threshold value. If the learning does not converge even if all the text data included in the text corpus to be learned is extracted, the acquisition unit 203 may extract the text data sequentially from the first text data again. That is, in the learning process, text data is cyclically extracted from the text corpus to be learned, and the values of the weight matrices W ^XH and W ^HY are converged.

  As described above, the semantic vector learning unit 204 receives the text data in the text corpus as a teacher signal and inputs a certain word, and the weight matrix of the neural network is set so that the appearance probability of the word before and after the word increases. To learn the semantic vector assigned to the word. For this reason, inevitably, in a plurality of words included in the text corpus, words having similar word strings before and after have similar semantic vectors to be learned. This is because learning based on the distribution hypothesis that words having the same meaning appear in similar contexts is performed.

  However, in an actual natural language text, antonyms often appear with a similar context. As can be seen in the examples of the text corpora 201C and 201D shown in FIG. 4, the contexts of the words having antonyms are often the same as or similar to each other. For this reason, when learning based on the distribution hypothesis is performed using a normally collected text corpus as learning data, the semantic vectors assigned to the words having the antonym relationship are similar, and it becomes difficult to clearly distinguish the two.

  FIG. 11 is a graph in which the semantic vectors assigned to the word “up” and the word “down” are reduced in two dimensions by the principal component analysis method in the semantic vector table of the comparative example of the present embodiment. The semantic vector table of this comparative example is obtained by performing learning based on the distribution hypothesis on a text corpus created by morphological analysis of the Japanese version of Wikipedia. As shown in FIG. 11, the word “up” and the word “down” are arranged at close positions, and it is understood that very close semantic vectors are assigned.

  FIG. 12 is a graph in which the semantic vectors assigned to the word “up” and the word “down” in the semantic vector table 207 in this embodiment are reduced to two dimensions by the principal component analysis method. In FIG. 12, a text corpus created from the Japanese version of Wikipedia is used as the general text corpus 201. Also, a text corpus created so that the antonym context shown in 202A of FIG. 5 differs for antonyms including the word “up” and the word “down” is used as the antonym text corpus 202. . Then, learning based on the distribution hypothesis is performed for both text corpora, and a semantic vector table 207 is created.

  That is, in the semantic vector table 207, the first vector representing the meaning of the first word and the first word are stored in association with each other, and the distance from the first vector in the vector space is a predetermined distance or more. The vector and the second word are stored in association with each other.

  As shown in FIG. 12, the word “up” and the word “down” are arranged far apart compared to FIG. 11, and it can be seen that a significantly different meaning vector is assigned.

  As described above, in the word meaning information generation apparatus according to the present embodiment, the antonym text corpus 202 created so that the contexts of the words having antonym relations are different is used in addition to the normal general text corpus 201. . Since learning based on the distribution hypothesis is performed for both text corpora, it is possible to assign a semantic vector to each word so that words having antonym relations can be appropriately distinguished from each other.

  As described above, the present disclosure has been described using an example of distinguishing a word from its antonym. In addition to this, the present disclosure can be applied to the following specific examples.

  (1) For example, the first text corpus 101 includes natural language text data used for an instruction to operate the device, and the second text corpus 102 includes a word related to the operation content of the device as a third word. can do. By doing this, for example, “Please raise the temperature”, “Please lower the temperature”, “Turn on the air conditioner in the bedroom” and “Turn on the air conditioner in the living room”. However, it is possible to appropriately distinguish instructions having different meanings and prevent erroneous operation of the device.

  (2) Also, the first text corpus 101 includes natural language text data used for explanation of symptoms by a patient in medical diagnosis, and the second text corpus 102 includes a word related to the physical condition as a third word. can do. By doing so, for example, appropriately distinguishing descriptions of symptoms that are similar in word sequence but completely different in meaning, for example, “My head hurts from three days ago” and “My head sways from three days ago” It is possible to prevent wrong diagnosis.

  (3) The first text corpus 101 includes natural language text data used for explanation or treatment of symptoms in medical diagnosis, and the second text corpus 102 includes words relating to body parts as third words. It can be constituted as follows. By doing this, for example, word strings like “right hand hurts from 3 days ago” and “abdomen hurts from 3 days ago” or “cool your head” and “cool your left foot” are similar. However, it is possible to appropriately distinguish between explanations of symptoms or treatments with completely different meanings, and prevent wrong diagnosis or presentation of wrong treatments.

  (4) Further, the first text corpus 101 includes natural language text data used for explanation of treatments for symptoms in medical diagnosis, and the second text corpus 102 is configured to include words related to treatment contents as third words. be able to. In this way, for example, “Please warm the affected area” and “Cool the affected area”, but appropriately distinguish between descriptions of treatments that have similar word strings but have completely different meanings, and present incorrect treatments. Can be prevented.

  Next, a usage form of the semantic information table created by the above learning will be described. FIG. 13 is an example of a block diagram of a household electrical appliance 300 that constitutes a first example of a usage form of the semantic information table.

  The home appliance 300 includes various home appliances such as a television, a recorder, an audio device, a washing machine, an air conditioner, a refrigerator, and a lighting device.

  The home appliance 300 includes a semantic information table 301, a microphone 302, an audio processing unit 303, an analysis unit 304, a command generation unit 305, and a command execution unit 306. The semantic information table 301 is a table created by performing the above learning on the first and second text corpuses 101 and 102 shown in FIG. 1, and corresponds to the semantic information table 107 in FIG.

  The microphone 302 converts the sound into an electric sound signal, and is used to collect the user's sound. The voice processing unit 303 analyzes the voice signal output from the microphone 302 and generates text data indicating the voice uttered by the user. The analysis unit 304 analyzes the text data generated by the voice processing unit 303 using the semantic information table 301. Here, the analysis unit 304 determines semantic information of each word constituting the input text data by referring to the semantic information table 301. Then, the analysis unit 304 determines whether or not the user's utterance content is related to an operation on the home appliance 300 from the determined semantic information of each word. If the user's utterance content relates to an operation on home appliance 300, analysis unit 304 outputs operation information indicating the operation to command generation unit 305.

  The command generation unit 305 generates a command for executing the operation indicated by the input operation information and outputs the command to the command execution unit 306. The command execution unit 306 executes the input command. As a result, the home appliance 300 can appropriately recognize the operation content that the user speaks using the semantic information table 301.

  FIG. 14 is an example of a block diagram of a home appliance system that constitutes a second example of the usage form of the semantic information table. In this home appliance system, a server 500 existing on the cloud has a voice recognition function, and the home appliance 400 is operated by voice.

  The home appliance 400 and the server 500 are connected via a public communication network such as the Internet, for example.

  The household electrical appliance 400 is the same as the household electrical appliance 300 demonstrated in FIG. However, in the second example, since the home appliance 400 does not perform voice recognition, the semantic information table 501, the voice processing unit 502, the analysis unit 503, and the command generation unit 504 are provided in the server 500.

  The microphone 401 is the same as the microphone 302 in FIG. The signal processing unit 402 determines whether or not the audio signal input from the microphone 401 is noise. If the audio signal is not noise, the signal processing unit 402 outputs the audio signal to the communication unit 404. The communication unit 404 converts the input audio signal into a communication signal having a communicable format and transmits it to the server 500.

  The communication unit 505 of the server 500 receives a communication signal from the home appliance 400, extracts an audio signal, and outputs the audio signal to the audio processing unit 502. Similar to the speech processing unit 303 shown in FIG. 13, the speech processing unit 502 analyzes the input speech signal and generates text data indicating the speech uttered by the user. Similar to the analysis unit 304 shown in FIG. 13, the analysis unit 503 analyzes the text data generated by the speech processing unit 502 using the semantic information table 501, and outputs operation information to the command generation unit 504.

  The command generation unit 504 generates a command for executing the operation indicated by the input operation information and outputs the command to the communication unit 505. The communication unit 505 converts the input command into a communication signal having a communicable format and transmits the communication signal to the home appliance 400.

  The communication unit 404 of the home appliance 400 receives a communication signal, removes a header or the like from the received communication signal, and outputs the header to the signal processing unit 402. When the communication signal from which the header has been removed is a command of home appliance 400, signal processing unit 402 outputs the command to command execution unit 403. The command execution unit 403 executes the input command.

  As described above, in the home appliance system of FIG. 14, the server 500 appropriately recognizes the operation content spoken by the user using the semantic information table 501 generated by the learning, and transmits a command to the home appliance 400. be able to.

  The word meaning information generating apparatus according to the present disclosure is effective when applied to an application that handles the meaning of natural language text. For example, the present invention can be applied to retrieval of sentences having similar meanings, paraphrase processing, semantic classification of uttered sentences in a dialogue system, and the like.

DESCRIPTION OF SYMBOLS 101 1st text corpus 102 2nd text corpus 103 Acquisition part 104 Semantic information learning part 105 Semantic information management part 106 Corpus generation part 107 Semantic information table 108 Operation part 110 Storage part 120 Processing part 130 Storage part 201 General text corpus 201A, 201B , 201C, 201D Text corpus 202 Antonym text corpus 202A, 202B Text corpus 203 Acquisition unit 204 Semantic vector learning unit 205 Semantic vector management unit 206 Corpus generation unit 207 Semantic vector table

Claims (16)

A semantic information generation method performed by a semantic information generation device,
Get text data,
Based on a semantic information table in which words and vector information representing the meaning of the words in a vector space of a predetermined dimension are associated, the meaning of the text data is analyzed,
Outputting information indicating the analyzed meaning;
The semantic information table is
Based on the arrangement of word strings in the first text corpus and the second text corpus, a first vector representing the meaning of the first word in the vector space is assigned to the first word, and the meaning of the second word in the vector space is assigned. Generated by learning to assign a second vector to the second word,
The first text corpus includes the first word, includes first text data of a first sentence written in a natural language, the second word whose meaning should be distinguished from the first word, and the second The second word distribution indicating the type and number of occurrences of the word appearing in the predetermined range before and after the word is similar to the first word distribution in the predetermined range before and after the first word in the first sentence. Second text data of the second sentence,
The second text corpus includes a third word that is the same word as at least one of the first word and the second word, and the third word distribution of the predetermined range before and after the third word is the first word. Including the third text data of the third sentence that is not similar to the one word distribution,
Semantic information generation method. The second text corpus includes the third word and an artificially created fourth word that does not appear in natural language text data;
In the third text data, the word included in the predetermined range before and after the third word is the fourth word,
The semantic information generation method according to claim 1. The first text data and the second text data are composed of words in a first language,
In the third text data, the third word is a word in the first language, and words included in the predetermined range before and after the third word are in a second language different from the first language. The semantic information generation method according to claim 1, which is a word. The semantic information generation method according to claim 1, wherein the second word is an antonym for the first word. The second word has the same meaning as the first word, and is a different word from the first word.
The semantic information generation method according to claim 1. The second word belongs to the same concept as the first word and has a different attribute from the first word.
The semantic information generation method according to claim 1. The learning is performed using a neural network.
The semantic information generation method according to claim 1. The learning is performed using latent semantic indexing,
The semantic information generation method according to claim 1. The learning is performed using probabilistic semantic indexing.
The semantic information generation method according to claim 1. The vector space has the number of different words appearing in the first text corpus and the second text corpus as the number of dimensions.
The semantic information generation method according to claim 1. The first text corpus includes natural language text data used for instructions to operate the device,
The third word is a word related to the operation content of the device.
The semantic information generation method according to claim 1. The first text corpus includes natural language text data used to explain symptoms by a patient in a medical diagnosis,
The third word is a word related to a physical condition.
The semantic information generation method according to claim 1. The first text corpus includes natural language text data used for explanation of symptoms or treatment for the symptoms in medical diagnosis,
The third word is a word related to a body part,
The semantic information generation method according to claim 1. The first text corpus includes natural language text data used to describe treatment of symptoms in medical diagnosis,
The third word is a word related to treatment content.
The semantic information generation method according to claim 1. Means for obtaining text data;
Means for analyzing the meaning of the text data based on a semantic information table in which a word and vector information representing the meaning of the word in a vector space of a predetermined dimension are associated;
Means for outputting information indicating the analyzed meaning,
The semantic information table is
Based on the arrangement of word strings in the first text corpus and the second text corpus, a first vector representing the meaning of the first word in the vector space is assigned to the first word, and the meaning of the second word in the vector space is assigned. Generated by learning to assign a second vector to the second word,
The first text corpus includes the first word, includes first text data of a first sentence written in a natural language, the second word whose meaning should be distinguished from the first word, and the second The second word distribution indicating the type and number of occurrences of the word appearing in the predetermined range before and after the word is similar to the first word distribution in the predetermined range before and after the first word in the first sentence. Second text data of the second sentence,
The second text corpus includes a third word that is the same word as at least one of the first word and the second word, and the third word distribution of the predetermined range before and after the third word is the first word. Including the third text data of the third sentence that is not similar to the one word distribution,
Semantic information generator. Get text data to computer,
Based on a semantic information table in which words and vector information representing the meaning of the words in a vector space of a predetermined dimension are associated, the meaning of the text data is analyzed,
A program for executing output of information indicating the analyzed meaning,
The semantic information table is
Based on the arrangement of word strings in the first text corpus and the second text corpus, a first vector representing the meaning of the first word in the vector space is assigned to the first word, and the meaning of the second word in the vector space is assigned. Generated by learning to assign a second vector to the second word,
The first text corpus includes the first word, includes first text data of a first sentence written in a natural language, the second word whose meaning should be distinguished from the first word, and the second The second word distribution indicating the type and number of occurrences of the word appearing in the predetermined range before and after the word is similar to the first word distribution in the predetermined range before and after the first word in the first sentence. Second text data of the second sentence,
The second text corpus includes a third word that is the same word as at least one of the first word and the second word, and the third word distribution of the predetermined range before and after the third word is the first word. Including the third text data of the third sentence that is not similar to the one word distribution,
program.

Images